TMA4245 Statistikk Eksamen 9. desember 2013

Transkript

1 Eksamen 9. desember 2013 Norges teknisk-naturvitenskapelige universitet Institutt for matematiske fag Løsningsskisse Oppgave 1 a) Define the following events: A: Getting an ace as your first card B: Getting a blackjack Since the suit of a card does not matter in this game and we have a situation with replacement, we can consider the deck to have 13 cards with just one ace, king, queen, jack and ten. Then, dividing the number of favorable by the number of possible hands, disregarding the order in which the cards are dealt, gives and P (B) = /2! = = , P (B A) = 4/13 = When P (A) = 0.1, P (B) = 0.06, and P (A B) = 0.4, we have that P (B A) = P (A B) P (B) P (A) = = b) This situation satisfies the conditions for a Bernoulli process: each game results in either a loss or a win for Lars, the probability of a win is constant, and the games are independent. The number of the trial in a Bernoulli process on which the first success occurs, follows a geometric distribution. Thus, the distribution of X is f(x) = g(x; p) = p(1 p) x 1 = 0.3 (0.7) x 1, where p = 0.3 is his chance of winning any one game. Lars s bet in a game is 2 x 1, where x is the number of the game i.e., 2 0 = 1 for the first game, 2 1 = 2 for the second game, 2 2 = 4 for the third game, 2 3 = 8 for the fourth game and 2 4 = 16 for the fifth game, if he has not woefore that. Since his winnings are twice his bet, W = 2 2 X 1 = 2 X. Therefore, W is a function of X and E(W ) = 2 x f(x) = p 2 x (1 p) x 1 = 0.3 ( ) = eksdes13-lsf-b 8. april 2016 Side 1

2 So Lars s expected winnings are 6 dollars and 57 cents. TMA4245 Statistikk There are two ways of finding E(Y ). The most elegant solution involves realizing that if Lars wins, his winnings will be one dollar more than his previous losses. E.g., that 4 (1 + 2) = 8 ( ) = 1, or in general: 2 i 1 = 2 n 1. (This formula is given in a more general form at the top of page 112 in Rottmann.) Then we see that E(Y ) = 1 P (Lars wins) 2 i 1 P (Lars loses 5 times) = (1 (1 p) 5 ) (2 5 1) (1 p) 5 = ( ) = 4.4. The more straightforward approach to finding E(Y ) is to compute the expected value of Lars s bets and subtract that from E(W ): x E(Y ) = E(W ) P (X = x) 2 j 1 P (X > 5) 2 j 1, where P (X = x) x 2 j 1 = 1 f(1) + (1 + 2) f(2) + ( )f(3) + ( )f(4) + ( )f(5) = p + 3p(1 p) + 7p(1 p) p(1 p) p(1 p) 4 = 0.3 ( ) = 5.73, and P (X > 5) = 1 P (X 5) = 1 p(1 p) x 1 = ( ) = = 0.17, and 4 2 j 1 = = = 31, so that E(Y ) = = 4.4. Thus, Lars can expect to walk out of the casino 4 dollar and 40 cents poorer. eksdes13-lsf-b 8. april 2016 Side 2

3 Oppgave 2 a) Expected value in the binomial distribution is E(X) = np = = 16. P (X > 16) = 1 P (X 16) = = P (X = 20 X > 16) = P (X=20) P (X>16) = = = b) H 0 : p = 0.8, H 1 : p > 0.8. Normal approximation means Z = X E(X) = Var(X) X np np(1 p) = X is assumed Gaussian. We reject H 0 when the observed fraction of success is significantly large. Setting α = 0.1 the rejection value is z 0.1 = We get X = 18 and Z = = Since Z < z 0.1, the observed number is not significantly large. We do not reject H 0. Half correction could also be used in this exercise, and gives more accurate results. In that case we get Z = = Full score is given to both methods. c) The hypothesis test is rejected when the observation is extreme. In particular, we reject if the observation is larger than c, and the critical level c is determined as the smallest c that satisfies P (X > c p = 0.8) α. For various c we get: P (X > 19 p = 0.8) = 0.011, P (X > 18 p = 0.8) = 0.07, P (X > 17 p = 0.8) = At significance level α = 0.1 we reject when X > 18. Based on the observed value X = 18, the P value is P (X 18 p = 0.8) = 1 P (X 17 p = 0.8) = The power is the probability of rejection, given that the parameter is H 1 : p = 0.9: P (reject p = 0.9) = P (X > 18 p = 0.9) = 1 P (X 18 p = 0.9) = = Oppgave 3 a) Variansen til utvalgsgjennomsnittet er ( ) ( Var( X) 1 ) = Var X i = 1 n n 2 Var X i = 1 n 2 Var(X i ) = 1 n 2 σ 2 = 1 n 2 nσ2 = σ2 n. Sannsynlighetstetthetsfunksjonen til normalfordelingen er gitt på s. 25 i Tabeller og formler i statistikk som f(x) = 1 ( exp 1 (x µ) 2 ) 2πσ 2 σ 2, slik at vi har f(µ) = 1 ( exp 1 ) 2πσ 2 0 σ 2 = 1 e 0 = 1. 2πσ 2πσ eksdes13-lsf-b 8. april 2016 Side 3

4 Dette gir at hvilket skulle vises. Var( X) = 1 4n ( f(µ) ) 2 = 1 ( ) 2 = πσ2 4n 1 2n = π Var( X), 2 2πσ Når vi skal velge mellom to estimatorer som begge er forventningsrette, velger vi alltid den med minst varians. Siden π > 1 har vi Var( X) > Var( X), som betyr at vi foretrekker å bruke X som estimator for µ. b) På grunn av de to tydelige outlierne på oppsiden, kommer medianen X til å være mindre enn utvalgsgjennomsnittet X (for disse dataene er X = mens X = 175.3). Vi har antatt at rekruttenes høyder er normalfordelte. Utfra histogrammet ser det ut til at gjennomsnittet ligger rundt 170 cm. I så fall er sannsynligheten for at to av de tretti datapunktene er større enn 235 cm neglisjerbar, så de ekstreme verdiene til disse to datapunktene skyldes antakelig en feil hos rekrutten som fylte inn dataene i regnearket ikke spesielt usannsynlig, gitt det gulnede papiret og falmede blekket. Siden utvalgsgjennomsnittet er følsomt for outliere, mens utvalgsmedianen ikke er det, gir medianen et bedre estimat enn gjennomsnittet i dette tilfellet. Anmerkning vedrørende dataene Datasettet i denne oppgaven er naturligvis fiktivt. Histogrammet er laget for 28 datapunkt trukket tilfeldig fra en normalfordeling med forventningsverdi 166 cm (litt lavere enn gjennomsnittshøyden for 1878, som er cm) og standardavvik 7 cm, og med to outliere på 239 cm og 251 cm (høyden til verdens høyeste mann). Når X N(166, 7 2 ) så er P (X 239) = Oppgave 4 a) ˆβ 1 = (x i x)(y i ȳ) (x i x) 2 = (x i x)y i (x i x)x i = = x iy i 1 n x i y i x2 i 1 n ( x i) 2 = x iy i x y i x2 i x x i = ˆβ 0 = 1 n y i ˆβ 1 1 n x i = = We have Var( ˆβ 1 ) = σ 2 /( (x i x) 2 ). Then T = ˆβ 1 β 1, where s b = s (x i x) 2 = s b s n. x2 i (1/n)( x i) 2 eksdes13-lsf-b 8. april 2016 Side 4

5 Based on the t-distribution P ( t n 2,α/2 < T < t n 2,α/2 ) = 1 α. This gives: P ( ˆβ 1 t n 2,α/2 s b < β 1 < ˆβ 1 + t n 2,α/2 s b ) = 1 α. Using t 22,0.025 = 2.07 we get: (115 ± / (1/24)(925) 2 ), = (115 ± 46) = (69, 161). b) Prediction at week 40: Ŷ 0 = ˆβ 0 + ˆβ 1 40 = = Ŷ 0 Y 0 is Gaussian distributed. Because of unbiased estimates ˆβ 0 and ˆβ 1, we have mean E(Ŷ0) = β 0 + β We also have E(Y 0 ) = β 0 + β This means E(Ŷ0 Y 0 ) = 0. We next use that Ŷ0 = ˆβ 0 + ˆβ 1 x 0 = Ȳ + ˆβ 1 (x 0 x). The mean of the data is x = (1/24)(925) = The variance is Var(Ŷ0 Y 0 ) = Var(Ŷ0) + Var(Y 0 ) = Var(Ȳ ) + (40 x)2 Var( ˆβ 1 ) + Var(Y 0 ) = σ 2 /n + σ 2 (40 x) 2 (x i x) 2 + σ2 We estimate σ 2 with s 2 and we get s 0 = s (40 x)2 + n n (x i x) 2 = ( ) (1/24)(925) 2 = 201 and get a statistic with a T-distribution: Ŷ 0 Y 0 s 0 t(n 2). Then, P ( t n 2,α/2 < Ŷ0 Y 0 s 0 < t n 2,α/2 ) = 1 α, and this gives: P (Ŷo t n 2,α/2 s 0 < Y 0 < Ŷ0 + t n 2,α/2 s 0 ) = 1 α. The significance level α = 0.1 means we set t 22,0.95 = This gives prediction interval 3135 ± (2 789, 3 481). The interval at week 42 has the same form, but now we have a term ( ) 2 instead of ( ) 2 in the variance. ( ) 2 This gives only a small increase in the width since = 0.16 and (1/24)(925) are both relatively small compared with 1. We get width at week 40: and width at week 42: (1/24) = 690, (1/24) = 732. The prediction interval is narrowest at the mean of the week data, i.e., at week 39. ( ) (1/24)(925) 2 = eksdes13-lsf-b 8. april 2016 Side 5

6 c) From the plot we clearly see that boys are heavier than girls on average. The increase with gestational weeks seem similar, and for this reason we fit the same slope. The error level seems to be about the same for boys and girls, and there is no reason to use a different variance. The model would predict a non-zero intercept term and different weight for boys and girls at 0 weeks, which is unntatural. Perhaps a model with different slopes and intercept 0 would be more appropriate. Then again the linear regression model is valid only within the region where data is available. SSE = (y i β b β 1 x i ) 2 + (y i β g β 1 x i ) 2 We differentiating this expression and set the derivative equal to 0. dβ g dβ b = = (y i β b β 1 x i ) = β b + β 1 x i + y i = 0 From these two first expressions we have: Here, n g = n = 12. (y i β g β 1 x i ) = n g β g + β 1 ˆβ b = (1/ ) y i ˆβ b (1/ ) ˆβ g = (1/n g ) y i ˆβ 1 (1/n g ) x i x i + x i We have to insert these into the expression for β 1 to solve for the slope: y i = 0 dβ 1 = x i (y i β b β 1 x i ) x i (y i β g β 1 x i ) dβ 1 = x i y i + β b x i + β g Now, inserting for the solution to the slopes we get: x i + β 1 x 2 i dβ 1 = a + β 1 b = 0, ˆβ 1 = a/b where a = x i y i (1/ ) y i x i (1/n g ) y i x i eksdes13-lsf-b 8. april 2016 Side 6

7 b = x 2 i (1/ )( x i ) 2 (1/n g )( x i ) 2 Inserting numbers we have: a = (1/12) (1/12) = 9 007, b = (1/12) (1/12) = 75. Finally, ˆβ 1 = 9 007/75 = ˆβ b = / (1/12) 460 = ˆβ g = / (1/12) 465 = eksdes13-lsf-b 8. april 2016 Side 7